Classifying Multistate DNA Origami: An Automated Approach with Minimal Labeling and Confidence-Based Filtering

DNA origami nanostructures can adopt multiple conformational states, making an accurate and rapid classification essential for advancing nanoscale fabrication. Although atomic force microscopy (AFM) provides indispensable structural validation, manual per-object labeling for classification is a persistent bottleneck. Here, we introduce an automated classification framework that minimizes user annotation while maintaining high accuracy. The pipeline employs a confidence-based filter to flag uncertain or aggregated objects, which are then selectively relabeled and augmented for iterative retraining. This human-in-the-loop refinement enables the classifier to capture greater intraclass morphological diversity without extensive new labeling. With only 20 labeled images per class expanded through strong augmentation, our framework consistently outperforms established few-shot and contrastive learning baselines on two structurally distinct DNA origami data sets, while also shortening analysis time relative to conventional methods, enabling efficient, high-throughput, and objective classification of multistate DNA nanostructure populations.